duggal ch habra

7/31/2019 Duggal Ch Habra

http://slidepdf.com/reader/full/duggal-ch-habra 1/7

179

LEARNING SYSTEMS AND THEIR APPLICATIONS:FUTURE OF STRATEGIC EXPERT SYSTEM

Dr. Sudesh M. Duggal, Northern Kentucky University, [email protected]

Mr. Rahul Chhabra, Northern Kentucky University

ABSTRACT

“Learning” denotes that a system is capable of adapting to a task, such that when the task is

repeatedly performed the system will perform the task more efficiently than on previous attempts.

Learning system may incorporate neural networks, fuzzy logic, genetic learning, self-learning

expert systems, or a combination of any or all of these technologies. These systems are capable

of the self-generation of mathematical, logical, or analytical rules for determining desired

outputs based upon some input criteria or user assistance. Some of these systems are self-

training; others require training and/or previous knowledge in order to learn. Learning systems

can be hardware or software-based, or a combination of both. Today, a great deal of research

incorporating machine-learning techniques continues in the area of Artificial Intelligence. The purpose of this paper is to discusses their strategic applications, advantages and problems of

learning systems, then focus on how these technologies are being used for developing solutions

to difficult, yet strategic decisions, and finally present recommendations for how future systems

will have to differ from current systems.

Keywords: Learning Systems, Neural Networks, Fuzzy Logic, Self-Learning expert Systems

LEARNING STARTEGIES

Learning Strategies will typically fall into one of the following five categories of learning

paradigms (1) Supervised Concept Learning, (2) Conceptual Clustering, (3) Analytical Learning,(4) Genetic Algorithms, and (5) Connectionist learning. (Derge, 1992) Supervised concept

learning is the most mature of the learning paradigm, and not surprisingly, has had the mostapplications to date. The primary difference between conceptual clustering and supervisedlearning is that conceptual clustering systems must recognize the similarities between the

training sets, and cluster these groups in accordance with a pre-established notion of similarity.Analytical learning attempts to speed the learning process by improving the goal searching

method. Genetic algorithms utilize adaptive search methods, based upon the Darwinian conceptsof “survival of the fittest.” Connectionist learning evolved out of the study of “perception,” andis a neural based learning paradigm utilizing massively parallel back propagation learning

algorithms.

LEARNING TECHNOLOGIES

Four state-of-the-art learning technologies are discussed in this paper. These are Neural

Networks, Fuzzy Logic, Genetic Learning, and Self-Learning Expert System. Each of the abovetechnologies are gaining rapid acceptance across the globe as more and more applications are

implemented using these technologies and are discussed below along with their operations,advantages, and disadvantages.



IACIS 2002 LEARNING SYSTEMS AND THEIR APPLICATIONS: FUTURE OF STRATEGIC …

180

Neural Networks

The development of neural networks was initiated decades ago, in an attempt to mimic theprocess of the human brain. Neural networks typically require tens, hundreds, or even thousands

of passes of the data sets before accurate classification can be accomplished. Once the network

is trained and “good” results are determined, the neural network can usually perform patternrecognition, classification, or generate predictions on data that the network has never “seen”

before. Table 1 given below, differentiate and exemplify the major differences between neuralnetworks and expert systems (Samdani, 1990).

Neural networks have an additionaladvantage over other learning

system, in that they are faulttolerant; neural nets can have node

failures and still generates “good”results. Table 1 above states that

neural nets need a database andexpert systems need a human expert.Self-learning expert systems will

make use of a database as well, andmany case-based expert systems willalso utilize a database. In these

instances, an expert system requiresan interface to data set storage. One may also have the ability to load applications specific pre-

trained neural net parameters into the neural net, eliminating having to repeat the trainingprocess. Also, self-learning expert systems will require a human expert, but perhaps only for theinitial design of the rule base, and then to supervise the process of the learning process. Neutral

networks have been used in hundreds of applications: assisting in process control applications,stock market forecasting, real estate sales forecasting, pattern recognition, and many other

diverse application areas.

Fuzzy Logic

Fuzzy logic employs mathematical relationships to determine a solution to a target problem.

However, unlike conventional solutions that yield one precise solution to a problem, fuzzy logicallows some tolerance 9or overlap) between possible solutions. Solutions to some problems aredefined as “membership groups" In operation; fuzzy logic is used in conjunction with a rule

based. Rules examine the input criteria, compare this to the desired criteria as determined by therules, and then determine the error that exists between input and the desired target. This may

result in a solution that is between membership functions. In these instances, the area under thecurve between the error function and the membership function is computed for both membershipfunctions. The centroid (center of gravity) is typically computed on the overlapping areas, and

this value is used for feedback into the system. Such a process is called “defuzzification”, andcalculation of the centroid is just one example of an algorithm used for accomplishing the

defuzzification process.

Table 1: Neural Nets Vs. Expert Systems

Neural Networks Expert Systems

Example Based Rule Based

Domain Free Domain Specific

Finds Rules Needs Rules

Little Programming Much Programming

Easy to Maintain Difficult to Maintain

Fault Tolerant Not Fault Tolerant

Needs a Database Needs a Human Expert

Fuzzy Logic Rigid Logic

Adaptive System Require Reprogramming



LEARNING SYSTEMS AND THEIR APPLICATIONS: FUTURE OF STRATEGIC … IACIS 2002

181

A classic problem solved by fuzzy logic controller problem of balancing the inverted Pendulum.One such solution uses a hardware implementation for defuzzification (a defuzzifier chip), and a

rule base-controlled (consisting of only 7 rules) fuzzy controller. Feedback from a positionsensor is fed to the fuzzy controller, where the correction signal is computed and fed to a DC

servomotor, which quickly corrects the position of the base containing the unbalanced pendulum.

Fuzzy systems, although becoming extremely popular in real-time control systems, have theadvantages of overcoming noisy data sets, but the disadvantage in being used as a broad

solution-solving tool, especially in those areas where tolerances are not permitted.

Genetic Learning

Genetic learning attempts to develop a model of a problem by applying possible solutions to a

target problem. A user-defined evaluator then determines how well the solution has “fitted” tothe target problem (note that the evaluator is an algorithm), using “survival of the fittest”

methods. That is, if a generated solution is not well fitted for the target, it is killed. If a solutionis a good fit, it survives. When a solution survives, it is then placed in a “gene pool,’ similar to

biological genetics, where other “good” solutions are placed. These solutions are then mutatedand/ or mixed with other “good” solutions in an attempt to create even better solutions.

Self-Learning Expert System

Most self-learning (or adaptive) expert systems are supervised learning systems. The most

common self- learning expert systems derive their rules based on a-priori information. Thesystem repeatedly guesses and improves upon its rule base, depending upon whether the guesses

were in the right or wrong direction. Eventually, the rule base is adjusted such that the logic of the inferences is able to handle all situations presented to the system. Thus, adaptive expertsystems alter their execution based upon past experiences. In operation, there are many methods

used to assist an expert system in adaptation. One could implement optimization methods fordecision tree adaptation (pruning), monitor histograms of rules and various inferences and

adjusted probabilities of decision search paths accordingly, etc.

The advantages of such a system are obvious – rules that have a higher frequency of occurring

are the first rules tried in a decision search. However, the disadvantages are glaringly obvious aswell: what if the system has some highly critical nodes that should always be searched first? It is

true that the system developer would hopefully remember to permanently keep such nodesenabled, but then the system is not fully optimized for searches. Self-learning systems may behighly recursive – if no goal is reached, then the existing success/failure ratios can be

continuously modified by the system in an attempt to reach a highly probable goal. There hasbeen a great deal of success with self-learning expert systems in many process control and

industrial applications, but these systems are typically found in conjunction with one or more of the aforementioned technologies such as neural networks, fuzzy logic systems or geneticalgorithms.

INTEGRATING LEARNING TECHNOLOGIES

Some of the most successful applications in recent research indicate that self-adaptive systemsare being implemented utilizing one or more learning technologies. This implies that expert




182

systems are being coupled with neural networks, fuzzy logic control, and genetic learning. Theadvantage of coupling several learning technologies together is obvious: by integrating several

learning technologies, one can create a system that maximizes its learning potential, capitalizingon the advantages of several different learning strategies. An example would be a system that

utilizes a neural network and a genetic learning algorithm. Recall that a neural network typically

converges rather slowly upon its solution, requiring several passes of training data sets in orderto determine the correct weights for the desired solution. The addition of the genetic learning

algorithm provides a more rapid convergence on the weights, which can be used by the neuralnetwork. An expert system could be added to determine when the genetic algorithm needs to be

invoked, adding an additional level of optimization. In this instance, the expert system coulddetermine the mean-squared-error between the desired output criteria of the neural network andthe actual output. Depending upon a tolerance or error threshold level, the expert system could

decide to call upon the genetic learning algorithm for the determination of better weights, whichcould then be used by the neural network, or allow the neural net to continue on its current

course of confluence. It would even be possible to add a fuzzy logic system into thisconfiguration to assist in determining the output error criteria, if the decisio n domain were

difficult to define or if the real-time control would be necessary.

One example of such a system is a product called NET-Link+ by Norrad (located in Nashua, N.

H.), which interfaces an expert system with a neural net, fuzzy logic, and genetic learning inorder to identify optimal telecommunication routes in real-time:

“You are no longer limited to a single (learning) paradigm. You can use neural networks forwhat they do best, fuzzy logic for what it does best, .. and genetic algorithms for what they

do best, all in the same application.” (Johnson, 1991)

Such highly integrated learning systems do not need to be extremely expensive, even for real-

time applications. It is possible to use a PC-based platform, interfaced with transputer boards(computers are high speed coprocessors that are on single cards which can be placed into the PC

back plane), for a real-time fully integrated learning system with capabilities of all learningsystems running in parallel including the expert system; each transputer board would run its ownapplication. Blackboard interfacing schemes (blackboards are commonly used with multiple

expert systems that need to access real-time data) could be utilized, and all system could accessthe data simultaneously.

Integrated learning systems could also be used in decision support systems for many predictionsand forecasting tasks. Highly integrated learning systems offer the best of all current state-of-

the-art learning paradigms, and should certainly be considered for automated decision system of the future.

APPLICATIONS OF LEARNING SYSTEMS

Last decade has seen incredible growth of Expert Systems incorporating intelligent algorithms.The penetration of Information Technology in business world has enabled technologies like

ANN, Fuzzy Logic and Genetic Learning to solve complex business problems. The latestapplications of ANNs and Fuzzy Logic are emerging in all industries including medicine and




183

chemistry. (Selén, 1998 and Virtanen, 1999) VTT Information Technology with participatingcompanies like Valio (Finland's largest dairy company), Kesko (Finland's largest wholesales

company) and ICL Data developed the Intelligent inventory control (VTT-PROMISE) system.The VTT-PROMISE project was concerned with the forecasting of product sales for companies

operating in the consumer market, by using an intelligent software tool. The goal was to prepare

the forecasts as automatically as possible, and in time spans consistent with the businessprocesses of the companies. The results seem promising but genetic algorithms haven't been

incorporated in the actual system due to sensitivity problems. A new method for the combinationof weekly and monthly forecasts has been developed (Karanta, 1999).

Another work conducted on similar lines was to utilize adaptive and intelligent methodsespecially for signal processing related applications. The group of companies involved was

Nokia Research Center (speech coding and recognition, image interpolation, bar-coderecognition), Nokia Mobile Phones (channel equalization), Oy Imix Ab (enhancement of digital

x-ray images) and Patria Finavitec Systems (passive detection of moving targets). The followingapplications were considered: speech recognition, speech coding, image interpolation, bar-code

recognition, channel equalization, enhancement of digital x-ray images and passive detection of moving targets. (Kantsila, 1999 and Salmela)

Another latest research is use of neural algorithms for radio communication systems. The novel-decoding algorithm for convolution codes has been derived at the Nokia Research Center. It usesa recurrent neural network, which has some similarities to the Hopfield network. The network is

dependent on the code structure and has to be derived for each code separately. The performanceof the codes has been verified using extensive computer simulations. The studies on detection in

a hostile environment concentrated on using self-organizing networks (SOMs). The main findingwas that neural methods might provide improved performance in detection problems wherenonlinearities are involved. (Raivio, 1997)

Genetic Algorithm is making its own niche in the intelligent systems world. In the past few years

GAs has shown their capability in advanced subjects like physics and economics. In the field of physics the important application is calculation of bound states and local density approximations.In Economics Genetic Algorithm has found its way in the Game theory (study of multi-person

decision problems). (Chughtai)

Fuzzy logic is also making its mark in the intelligent systems domain. Few applications whichhave made difference are Automatic control of dam gates for hydroelectric-power plants,Simplified control of robots, Substitution of an expert for the assessment of stock exchange

activities, Efficient and stable control of car-engines, Cruise-control for automobiles, Improvedefficiency and optimized function of industrial control applications, Recognition of handwritten

symbols with pocket computers, Controlling of subway systems in order to improve drivingcomfort, precision of halting and power economy and Improved safety for nuclear reactors. Oneof the most successful fuzzy logic implementations is the control of subway in Sendai, Japan.

The fuzzy system controls acceleration, deceleration, and breaking of the train. Since itsintroduction, it not only reduced energy consumption by 10%, but the passengers hardly notices

now when the train is changing its velocity. In the past neither conventional, nor human controlcould have achieved such performance (Soylemezogolu).




184

Fuzzy molecular modeling (FMM) is the application of fuzzy logic to molecular modeling.

Typically, to model a chemical's structure, lattice parameters and a coordinates table (such as theWyckoff coordinates for a chemical compound under examination) are needed. Together, the

lattice parameters and Wyckoff coordinates provide enough information to create a three-

dimensional (3-D) structure of a compound's unit cell. However, the measurement of latticeparameters is subject to variability from one experiment to another. The fuzzy lattice parameters

are created by searching the literature for experiments performed under the same conditions andthen extracting the minimum, average, and maximum values for each lattice parameter. The end

result is a fuzzy unit cell, with fuzzy bond lengths and fuzzy bond angles, that incorporates thevariability found in literature. However, two interesting phenomena arise in creating fuzzy unitcells, the concept of fuzzy lines and fuzzy vertices. By using fuzzy lattice parameters, it is

possible to express other characteristics of the unit cell in imprecise terms, thereby providingresearchers with possible ranges for characteristics of interest. This has the added benefit of

providing variable information that can help simplify a decision of selecting a specific compoundout of a selection of many. [Ress , Wyckoff, 1993]

The opening of Finland’s National Electricity Stock Market (EL-EX) a few years ago madepossible for electricity companies to trade their own production easily with other companies and

large-scale consumers. Prediction of future consumption in few next day scales can give a greatadvantage in both purchasing and selling electricity. To make this forecasting possible, anExpert System project was conceived. The aim of this project was to study the possibility of

predicting consumption in particular distribution area using neural networks. A further aim wasto design PC-software based on neural networks to help in the calculation and simulation of

different conditions. The results so far achieved were: - analysis of different neural networksenabled the identification of the optimal network design. However, the difference between theresults and realization must be quite small (less than 5 percentage units). The best model so far

gave difference of between 0 to 8 % and an average difference of 4.5 %. Considering thatconsumption in Finland varies very much, result was satisfactory (Laitinen).

LEARNING SYSTEMS OF FUTURE

Learning systems of the future must incorporate and take advantage of all paradigms. Expertsystem will certainly be the key to success for future self-adaptive systems. A decision support

system that allows for the integration of multiple expert systems, multiple learning paradigms,relational database technologies, and support modules that include natural language processing,hyper-technologies, statistical support, reporting modules, object-oriented code development,

object linking and embedding, dynamic link libraries, 4GL’s, etc. will be essential supportfunctions of the future systems. Future systems must be capable of allowing additional

applications-specific expert systems to be easily integrated and provide utilities for thedevelopment of expert systems. The expert system manager (EMS) would provide an intelligentfront-end to the decision support system. It would be capable of monitoring the execution of

several applications-specific expert systems (ASES), and have black boarding capabilities forreal-time data evaluation and sharing. The ASES’s would be adaptive systems, and have a

multitude of low and high level adaptive learning algorithms at their disposal.




185

Automated knowledge acquisition, a system’s capability of extracting knowledge from a userand generating the associated data and rules, will be another feature that learning systems of the

future will include. This feature would streamline expert system development and allow“unknown” data types, attributes, or objects to be realized by the system via the user’s input, or

by searching through existing databases. All levels of organizations would use future adaptive

systems, since their flexibility will allow users to communicate, share data, and solve problemsin the most efficient way possible. Intractable problems, unforeseen trends, new methods of

Research and Development, and more knowledgeable approaches to decision making are only afew of the benefits that will be realized from future adaptive systems.

REFERENCES

Chughtai (1995). Determining Economic Equilibria using Genetic Algorithms, by MelihaChughtai, in September 1995. Unpublished Thesis.

Coltman, John. “Is Artificial Intelligence Better Than the Real Thing?”, R&D Magazine, Sept.1990, Pg 64-68.

Johnson, C. “New Models Learn Genetically”, Electronic Engineering Times, Sept. 16, 1991, Pg.41-42.Kantsila A., Lehtokangas M. and Saarinen J., "Burst Adaptive Equalization of Binary Data,"

Journal of Intelligent Systems, vol. 9, no. 2, 1999.Karanta Ilkka: Multilevel forecasting improves corporate planning and operations. ERCIM

News, No. 38 (July 1999), p. 36.

Karanta Ilkka: Constrained forecasting with time series models. Bulletin of the InternationalStatistical Institute, 52nd session, contributed papers, Vol. 2, pp. 117-118.

Kirrane, Diane. “Machine learning”, Training & Development Journal,, 1990. Pg. 24-29.Laitinen Tommi HALT Ohjelmointi Oy. http://www.cis.hut.fi/neuronet/Tekes/11.shtml#21Raivio Kimmo, Jukka Henriksson and Olli Simula: Neural Detection of QAM signal with

strongly nonlinear receiver. In Proceedings of the WSOM, pp. 20-25, Espoo, Finland, June 4– 6, 1997

Ress David A., Using Fuzzy Logic for Molecular Modelinghttp://www.tms.org/pubs/journals/JOM/9908/Ress/Ress-9908.html

Salmela P., M. Lehtokangas and J. Saarinen, "Neural Network based Digit Recognition System

for Voice Dialling in Noisy Environments," International Journal of Information Sciences (toappear).

Samdani, G. “Neural Nets – They Learn From Examples”, Chemical Engineering, August 1990,Pg. 37-45.

Soylemezogolu Nazim The Logic of Fuzziness,

(http://www.math.harvard.edu/~hmb/issue2.1/FUZZY/fuzzy.html)Virtanen A, Gomari M, Kranse R, Stenman U-H. Estimation of prostate cancer probability by

logistic regression: free and total prostate-specific antigen, digital rectal examination, andheredity are significant variables. Clin Chem 1999; 45:987-994.

Wyckoff W.G., Crystal Structure (New York: Wiley-Interscience Publishers, 1963).

duggal ch habra

Documents